Online Feature Weighting for Human Discrimination in a Person Following Robot - Foundations on Natural and Artificial Computation - page 222

conditions such as dynamic, crowded and not structured environments. Our con-
troller must achieve a high performance even under changing light conditions,or
when there are people moving around and close to the person being followed
(wewillusetheword target to refer to this person henceforth). Finally, the
movement of the robot must be also safe, in the sense that it avoids colliding
with other people or with obstacles placed in the environment [1].
The majority of the solutions previously published choose to track the target
using either visual information[2] or data provided by a laser range scanner[1],
although the most robust solutions use a combination of both sensors[3,4].
In this paper, we developed a controller which combines both sensors to
achieve a person-following behaviour that fulfils two important characteristics:
adaptability and robustness. Our controller will not have to be modified, or the
parameters of the controller will not have to be tuned, to get the desired be-
haviour in different environments, or when the light conditions vary significantly.
The tracking of the target is done using data from the laser, easy to get and pro-
cess, and which offers a wide field of view, nevertheless the camera plays an
irreplaceable role to discriminate amongst the target and other people present
in the same scene (from now on we will call distractors to these people). Hence,
the data provided by the camera will correct or help the laser tracker whenever
the robot doesn't see person being followed for a short while, or when there are
distractors around the target.
2SyemOrvew
Fig. 1 shows an overview of our system. In this figure, we can see that our
system takes as inputs the images from a camera placed on the robot and data
coming from a laser scanner. Our system will provide the relative position of the
target with respect to the robot. We can notice the existence of two modules
clearly differentiated: the laser tracking module ,and the camera discrimination
module . The first one uses the leg detection algorithm developed by Gockley et
al. [1] to track the target. This algorithm combines the laser information with
the last known position of the target. The camera discrimination module works
on two clearly separated stages: first it uses the human detector presented by
Dalal et al. [5], to extract the torsos of the people present in the scene. These
torsos are the inputs of a target discrimination algorithm (second stage), which
compares a model of the target with each torso in the image. As a result of this
comparison each torso gets a dissimilarity measure. As we will describe later
in this paper, this dissimilarity measure is achieved after an adaptive weighting
process, that dynamically adjusts the importance of the different features used
to distinguish the distractors from the target. The model of the target that is
used to compare with the torsos detected in the image, is built during the first
detection of the target, and it is dynamically adjusted while the robot moves.
Whenever the camera module detects the target provides its position to the laser
module which will check and correct its own estimation of the target's position.
Finally, the position of the target is sent to a controller which, based on the